Preserving historical buildings is challenging due to limited resources and the inefficiency of manual inspections, especially for complex surfaces like masonry. Neural network-based crack detection methods have emerged, but they rely on large, diverse datasets, which are hard to collect and annotate for aging structures. This paper presents a new framework for generating synthetic crack datasets via 3D rendering, tailored to historic masonry. It combines real and synthetic data to train models that match the performance of those trained solely on real images. A novel crack simulation algorithm generates realistic patterns and annotations. Tests show the approach maintains segmentation accuracy while reducing the need for manual dataset creation.